Machine learning predicts functional outcome of DME

Machine learning models more accurately predicted best-corrected visual acuity (BCVA) outcomes at 4 and 24 weeks in eyes with diabetic macular edema (DME) in the phase 3 YOSEMITE and RHINE trials; One-year BCVA prediction remained complex.

This analysis, presented at the 2024 Association for Research in Vision and Ophthalmology (ARVO) Conference, examines the performance of machine learning models that use baseline features to predict functional outcomes over different treatment periods (short-term, medium-term, and long-term). It was centered around. His response to DME with faricimab, a bispecific antibody that blocks both VEGF-A and Ang-2 molecules.

In an interview with HCP LiveDaniela Ferrara, MD, chief medical director of Genentech's Ophthalmic Imaging Program, explained her thoughts on potential performance metrics for machine learning models, particularly the promising accuracy of the first and sixth visual function responses. Lunar time point.

“I think the important thing for clinicians like me is that we know that predicting visual function is a clinical challenge,” Ferrara told HCPLive. “In other words, when you have a patient sitting in front of you in the doctor's office or a patient enrolled in a clinical trial, you can't predict what their vision will be over time based solely on clinical considerations. It's very difficult to predict, and I'm very happy that some models were able to make such predictions.”

Patients receiving faricimab 6.0 mg every 8 weeks in Yosemite and Line were pooled and divided into a 70% development set, 15% trial set, and 15% holdout set. The development set was divided into five parts for cross-validation purposes. Target functional outcomes consisted of BCVA letter scores, and target time points were 4 weeks, 24 weeks, and 1 year.

Only baseline characteristics were assessed to predict BCVA over time. ElasticNet, Random Forest, Support Vector Machine, and eXtreme Gradient Boosting trees were trained and evaluated on the dataset. The model and input features that performed best on the test set were selected for evaluation on the holdout set for each target time point.

After analysis, based on the test results, the Random Forest model with Tier-1 input was selected to predict week 4, and the ElasticNet model with Tier-1A input was selected to predict week 24 and year 1. The researchers evaluated the model's performance using the percentage increase from the root mean square error (RMSE) of the test set to the holdout set.

The RMSEs of the holdout set for week 4, week 24, and year 1 forecasts were 8.26 (29%), 7.80 (22%), and 13.15 (99%), respectively. Within the holdout set, the Spearman correlations of residuals at the three study time points were statistically significant for each pair (P <.0001).

Ferrara said statistically significant correlations in the residuals indicate that the model's capabilities could be improved, either by larger datasets, new input features, or more advanced modeling. Therefore, he suggested that the performance gap may improve in the future.

“From a technical perspective, what we're thinking about next steps is to continue to publish baseline features and look at the results of this particular project to better mine clinical features at baseline. or how to interpret it,” Ferrara said.Said HCP Live.

“From a modeling perspective, I'm a clinician and not a data science expert, but we're going to continue to work in the very rapidly evolving field of artificial intelligence and machine learning. “Our group is at the forefront of developing new modeling approaches and evolving the work on the models themselves,” she added.

Disclosure: Daniela Ferrara is employed by Genentech, Inc.

References

_{Kikuchi Y, Abderezaei J, McLeod, Chen C, Benech AC, Ferrara D, Anegondi N, Yang Q. Prediction of functional outcome at different treatment durations with faricimab in diabetic macular edema (DME). Poster presented at the Association for Research in Vision and Ophthalmology (ARVO) 2024 Conference, May 5-9, 2024.}